Anomaly Explanation

Up till now we have made the decision about the class membership of a new activity instance, and the decision about it normality or abnormality. In case we decide that the new class member is anomalous, we need to explain the ways in which it is anomalous. We propose to achieve this objective based on the comparison of a new anomalous class member with respect to the typical class member of the selected class.

This idea is very similar to how humans try to understand new pieces of information - i.e. by comparing them with the mental models of the classes that they have developed in their minds. Since the model of a class in our case is the typical member of the class, we can attempt to "explain away" why something was labeled as anomalous member. We must mention here that the word explanation is being used in terms of its most naive nuance, i.e. just "finding" how one thing is different than another based on some "features". While our attempt to explain an anomaly is inspired by how it is usually theorized as to how humans explain something unusual, we do not claim to perform "explanation" in any amount of sophistication which the humans do. However, this idea of explanation, while very crude compared to the kind of explanation which humans can perform, can still be useful in applications such as Visual Surveillance since it would provide at least the first level of "explanation" to the user of the system about "what went wrong" when it went wrong. The human can then further analyze the anomaly to make more sense out of it, perhaps being partially influenced by the explanation that our system generated for her.

 

Recap: In this section we have described a way of "explaining away" in what ways something being detected as anomalous, was different from regular.